Abstract: The explosive growth of short-video platforms has aggravated the spread of fake news, calling for efficient automated detection methods. Existing methods predominantly rely on cross-modal interactions to capture semantic consistency; however, they are susceptible to deliberately fabricated content that achieves high inter-modal alignment, and they tend to ignore the reliability fluctuation of modal signals in complex environments, which severely limits their detection performance. To address these issues, a reliability-aware multi-view consistency network, termed RACH-Net, is proposed for short-video fake news detection. RACH-Net excavates deep logical conflicts and physical editing traces in videos from three heterogeneous views: the cross-modal consistency view (CMCV) that measures semantic agreement across modalities, the semantic-emotion view (SEV) that exposes the inconsistency between content semantics and expressed sentiment, and the spatio-temporal pattern view (STPV) that captures spatial–temporal cues and editing artifacts. Moreover, a cascaded reliability-aware perception module (CECP) is designed, which leverages evidence theory to quantify the quality of each unimodal signal and to detect cross-modal semantic conflicts, thereby shifting the fusion paradigm from similarity-driven to reliability-driven. A consistency-guided adaptive decision mechanism is further introduced to dynamically adjust the weights of the three views, endowing the model with self-correction capability. Experiments on the FakeSV and FakeTT datasets validate the superiority of RACH-Net, which achieves accuracies of 88.93% and 83.28%, respectively, and attains competitive detection performance.